CUDA · LLM Inference · AI Infrastructure

Building systems projects from first principles.

ACM silver medalist working on ContextOS, C++/CUDA, transformer inference, GPU kernels, KV cache, and small AI-infra tools that make complex systems easier to inspect and understand.

View GPT2.cu Read the write-up

$ ./build/gpt2_main --generate 8 --text "Hello, I am"
input text: Hello, I am
generated text: Hello, I am a little bit of a fan of the

Featured Work

Projects

C++ CUDA LLM Inference

GPT2.cu

Handwritten C++/CUDA GPT-2 inference with real Hugging Face weights, greedy generation, prefill/decode, and KV cache.

Speedup: 2.536x
Path: KV cache
Runtime: FP32

Repository Chinese write-up

Python GitHub Issues LLM Triage

IssueRadar

A CLI-first MVP that watches GitHub issues, applies deterministic filters, asks an LLM for fit analysis, and renders a daily digest.

Repository

Agent Context Workflow

ContextOS

An experimental personal context system for organizing project memory, technical notes, and agent workflows without turning them into a public diary.

Repository

Writing

Technical Notes

Zhihu 从零手写 CUDA GPT-2 推理：从 full recompute 到 KV cache

A long-form write-up about building GPT-2 inference in C++/CUDA, loading Hugging Face weights, and moving from naive generation to KV cache.

Focus